ISSN 2347 - 3983

Volume 9. No.11, November 2021

International
Mohammad M. Al-Momani Journal
et al., International Journal ofof Emerging
Emerging Trends
Trends in in Engineering
Engineering Research
Research, 9(11), November 2021, 1373 – 1378
Available Online at http://www.warse.org/IJETER/static/pdf/file/ijeter029112021.pdf
https://doi.org/10.30534/ijeter/2021/029112021

Prediction of Transient Stability Using Wide Area

Measurements Based on ANN
Mohammad M. Al-Momani1, Seba F. Al-Gharaibeh2
1
Protection Engineer, Jordan, monqedmohammad@gmail.com
2
Research Scholar, Jordan, 620180441015@muath.edu.jo

Received Date : October 02, 2021 Accepted Date : October 24, 2021 Published Date : November 07, 2021

ABSTRACT If an unstable case is predicted early, the prevention system

The concept of wide-area control and protection as an
application on real-time wide-area measurement systems
makes the transient stability prediction more accurate in
early time after fault occurrences. The transient prediction is
the first step in the dynamic control system to avoid any
unwanted emergency or non-stable power system state. In
this paper, an early predictionof the power system stability
once the fault cleaning using real-time dynamic data
collected by WAMS is proposed based on an artificial neural
network (ANN). The dataset collected by the different
contingency analyses on the IEEE 39 bus test system is used
to train a multilayer perceptron network. Pre-fault, during-
fault, and post-fault generators' speeds are fed to ANN as
inputs, and the status of the overall system, either stable or
not, is the output of ANN. The proposed model can predict
an unstable state within 100 ms after the fault. NEPLAN
simulator is used to simulate the dynamic analysis ofthe
IEEE 39-Bus test system, and MATLAB 2019a is used to
design the ANN.
Key words: wide-area monitoring system (WAMS), phasor
measuring unit PMU, transient stability, wide-area control
system, ANN, NEPLAN, MATLAB.
1. INTRODUCTION
Electric power systems should supply a continuous and
secure service for the consumers while considering the
economic dispatch. However, different types of faults
endanger their stability and, in the worst case, causing a loss
of stability, leading to a blackout of the power system [1-2].
Wide-area protection and control (WAPAC) systems are
proposed to prevent the power system loss of stability after
critical events [3-4]. WAPAC systems use the phasor
measurement units (PMU) information to counteract the
propagation of severe faults [5]. Predicting the power
system's post-fault transient stability afteran event is
significant for preventing any potential loss of stability. This
prediction increases the chance for the WAPAC systems to
trigger corrective control actions as fast as possible against
the potential instability, and hence, the dynamic security of
the power system is enhanced.
1373
from unwanted scenarios may be more applicable. The
WAMS plays a vital role in sensing, collecting, and
processing real-time power systems dynamically based on
GPS technology. Table 1: Comparative Analysis
The problem of power system transient instability prediction
after fault clearance is addressed in many articles. Based on
the literature, the problem may be categorized into (a) time-
domain simulations [6-7]. (b) transient-energy-function
(TEF) methods [8-9] and (c) machine-learning techniques
[10]. The time-domain simulation method is a dynamic
behavior of the power system, so a huge set of differential
equations need to solve. TEF method uses the Lyapunov
stability principle to determine the dynamic status of the
power system without using the complex differential
equations sets. ANN is a common example of machine-
learning techniques. All these techniques need to know the
operational point before the fault and the post fault status.
However, the power system's stability depends not only on
the post and pre-fault system operation but also on the fault
location, duration, and characteristic.
In [11], an artificial neural networks-based methodology is
proposed for directly predicting power system stability after
clearing the fault. A dataset is generated to train a multilayer
perceptron offline; then, it is used for early online prediction
of any transient instability. The inputs feed the neural
network: pre-fault, during-fault, and post-fault voltage
magnitude measurements collected from the phasor
measurement units. This study depends on the fault location
and duration information in addition to synchrophasor
measurement. However, this research provides a solution in
case the location and duration are well known.
Paper [12] proposes an approach for early detection of the
transient instability of the power system to initiate
emergency control in time. The synchrophasor
measurements are used for real-time monitoring of the
system. After that, the Artificial Neural Network (ANN) is
used as a classifier for predicting the transient instability
status of the system with rotor angles and speeds (frequency)
of the generator as inputs at different consecutive cycle
lengths after fault clearing. The stability status obtained from
ANN can initiate emergency control actions within a few
cycles from fault clearing.
Mohammad M. Al-Momani et al., International Journal of Emerging Trends in Engineering Research, 9(11), November 2021, 1373 – 1378
For this work, the ability of the ANN technique to build a
highly complex relationship between variables to estimate
the internal behavior and response characteristic of the
dynamic system, the capacity of ANN to work online with a
swift response, and the benefits of its accuracy in addition to
its ability to learn more cases in its live state are some
motivations to use ANN in a stability prediction. Data
collection from NEPLAN software for the IEEE 39-bus test
system and ANN structure are presented in section 2. Then,
the model simulation and ANN results are shown in section
3.
pre fault
during fault Y (iii) at the midpoint of each transmission line. The system
post fault
[
[
Input Output
Hidden
Figure 1. The general structure of ANN, the input layer (red
circles), and the output layer (green) depend on the number of input
and output. More than one layer may be used in the hidden layer
(yellow).
2. ARTIFICIAL NEURAL NETWORK ANN
Artificial Neural networks (ANN) or connectionist systems
are computing systems vaguely inspired by the biological
neural networks that constitute animal brains [13]. Such
systems" learn" to perform tasks by considering examples,
generally without being programmed with task-specific
rules.
ANNs have evolved into a broad family of techniques that
have advanced state-of-the-art across multiple domains. The
simplest types have one or more static components,
including the number of units, layers, unit weights, and
topology. On the other hand, dynamic types allow one or
more of these to evolve via learning. These types are much
more complicated but can shorten learning periods and
produce better results. Some types allow/require learning to
be supervised by the operator, while others operate
independently. Some types operate purely in hardware, while
others are purely software and run-on general-purpose
computers.
The mapping between input and targetbases on the concept
of neural where each neuron has its transfer function; there
are many presented transfer functions in the literature. Figure
1 shows the relation between the variables in our example.
For each generator, three angles are used, pre-fault, during-
fault, and post-fault.
1374
As mention above, ANN can be used to synthesize complex
mapping accurately and rapidly for practical applications.
Feedforward Neural Network (FFNN) is a class of ANN,
generally termed Multi-Layer Perceptron (MLP), maps input
data sets to a suitable output. The steps involved in detecting
transient stability following significant disturbance using
FFNN are presented in the next subsections.
2.1. Data collection
For transient stability assessment,a large number of
contingencies should be simulated. The selection of the
contingency depends on the operator's experience. In this
study, contingencies are considered at three different
locations: three-phase to ground fault (i) at the high side of
all the generating buses, (ii) at the critical load buses, and
should be simulated for all possible contingencies with
varying operating conditions to find system behavior after
disturbance for all possible operating scenarios. NEPLAN
generates considerable input data for three different loading
levels: base case and ±10% of the base case. Three-phase
faults have been applied at each substation, transmission line,
generator, and transformer and then removed after time.
Fault clearing time is randomly chosen from 5 cycles to 12
cycles. Long-duration faults are deliberately taken to
generate enough unstable cases. Thus, total operating points
corresponding to 300 cases are generated where 152 are
stable cases, and 148 are unstable cases.
2.2. feature selection
Three input features are extracted for each fault case:all
generators' pre-fault, during-fault, and post fault rotor speed.
Because all generators' pre-fault speeds should be the same
at each scenario for the same operation point, they can be
replaced by just two generators' speeds. These speeds
represent the operation point, so they are changed if the
operation point is changed.One generator angle is selected as
a reference angle, so it is ignored for the data input. So, the
ANN structure has 20 inputs and a single output.
3. SIMULATION AND RESULT
The proposed ANN-based approach is tested on the IEEE-39
New England test system. However,this system consists of
ten generators, forty-six transmission lines, and twelve
transformers. Moreover,the slack generator bus is number
39, andthe detail generator model is modeled with exciter,
turbines, stabilizers, and governors. The 3-phase fault is
created at all the high sides of generator buses, load buses,
and the midpoint of all the transmission lines at the base
case. The data is generated offline by randomly varying load
from 10 % of the base case for training of FFNN. The
dynamic simulation is performed for 5 seconds, and each
operating condition is assigned a class label of 0 or 1 for
transient unstable and transient stable state, respectively,
Mohammad M. Al-Momani et al., International Journal of Emerging Trends in Engineering Research, 9(11), November 2021, 1373 – 1378

based on the final values of rotor angles. The database of 300

cases is prepared to detect abnormal conditions, of which
152 are stable cases, and 148 are unstable cases. After
normalization, 70% of the scenarios are taken for training the

Pu
FFNN and the remaining 30% for testing and validation. All
the simulations are carried out using MATLAB 2019a, and
NEPLAN 5.5.5 is used to simulate IEEE 39 bus system,
shown in Figure 2. Two fault duration times are applied here,
short fault duration, 100 ms, and long fault duration,200 ms.
The effect of fault duration is shown in Figures 3 and 4.
These Figures show the dynamic change in generators angles
if a 3-phase fault occurs at bus 34, then the fault cleared after
100 ms and 200 ms in Figures 3 and 4, respectively. The
output of the trained ANN is also presented in these figures.
For the stable event, the output of the ANN is not changed
(zero) and will toggle for an unstable event. AVR-
Figure 4. Generators' speeds for a stable event, three-phase fault
applied at bus number 34 for 100 ms.
G0

G1 AVR-

N3

For ANN training, 300 samples, 20 inputs, and one output

N3
TR-

L2
L-2 6-
L-2 6-

L-2 8-
L data representing the most critical cases and extracting the
TR-

N2 N2

L-2 5-
N2 N2
best feature from the system are generated from NEPLAN to
L-0 2- L-2 6-

N0
L2

N2
L2 TR-

train ANN using MATLAB.

L-0 1- L0

L-0 2-
L-1 7-
L-1 7- L2
38
N

G0
AVR-
Two separate ANN models are used: the first is feed-forward
ANN, and the second isa step function preceptor. All
N0 N1 N1 N2

G0
AVR-
L-0 3- L1 L1 L-1 6- L-1 6-
L01

N0 N1 N3 combinations of three layers with less than 25 neurons per

L-0 1-
L0

N1
L-1 5-
L1
L-1 6-

L-2 1-
TR-
layer are trained, and the best system is selected based on the
G AVR-
L-0 3-

L-1 4-
L1
N2 N2
following criteria:
L L-0 4- L
L
L-2 3-

N1
N0 L-1 6-

∑( )
L3
L-0 4- L1 L-2 2-
= × 100% (1)
N0

L L L L
L-0 5- L1 L-1 3-

N0 N1 N1 N2
The minimum error structure is a three-layer with 5,5,8
L-0 9-
L-0 5-
L-0 6- L-0 6- TR- TR-
L TR- TR- neurons per layer, respectively; the structure is shown in
N0 TR- N1 N1 N2
TR-
N3
Figure 5. The system's error reaches 6.8e-12%. After
digitalize stage, this error is eliminated, which means the
L1 AVR-
L0 L1
L-0 7- L-1 0- L-1 0- TR-
G0
L1

N0 N3 N1
N3 N3 proposed ANN detects all stable and unstable scenarios
L-0 8-
L0
AVR-
L3
TR-
AVR-

G0 G0
AVR-
correctly.
G0

N3
N0

AVR-
L0
G0

Figure 2. IEEE 39 bus test system from NEPLAN simulator

consists of 10 generators and 46 lines.

Figure 5. The first ANN structure, FFNN, consists of three hidden

layers with 5,5,8 neurons per layer.
Pu

The backpropagation training algorithm conducts to

determine the optimal value of each weight in the structure.
The backpropagation method firstly selects random values to
the weights and derive the partial derivative of the error with
respect to each weight to minimize the error; therefore, this
technique is highly affected by the first random values,
which may (in most cases) lead the system to a local
solution. for overcoming this problem, the same structure [5
Figure 3. Generators' speeds for an unstable event, three-phase 5 8] is trained 2000 times. Then, the best solution defined by
fault applied at bus number 34 for 200 ms. equation (1) is saved.

1375
Mohammad M. Al-Momani et al., International Journal of Emerging Trends in Engineering Research, 9(11), November 2021, 1373 – 1378

The relative values of the weights between layersare

presented in Figure 6. The red square refers to a negative
value, green a positive value, and the big one represents
relatively considerable weight. The absolute values of these
weights are attached in Tables 1-4 in the appendix.
From Figure 6, the relative weights detect the importance of
the connection; for example, inputs number 2 and 12 are the
less important in this classification, but input number 13 is
the most important one. So, the input can be reselection
based on this character. Moreover, the location of PMUs can
also be selected based on the critical input (s). The ANN
structure is acceptable because the weights' values are
distributed uniformly between the neurons; so, all neurons
are essential to achieve the goal.

Figure 7.The performance plot of the trained ANN is divided into

70%, 15%, and 15% for train, test, and validation, respectively.

(a) Input Layer-Layer 1 relative wights

(b) Layer1-Layer 2 wights (c) Layer 2-Layer 3 wights

Figure 8.The confusion matrix of the trained ANN consists of 148

unstable events and 152 stable events; zero refers to a stable event
and one to an unstable event.

(d) Layer 3-output layer wights
Figure 6. Relative weights between input, output, and hidden
layers, green refers to a positive value, and red refers to a negative
value.
The performance plot is shown in Figure 7. The best
validation performance (cross-Entropy of the validation set)
is about 1.18e-7 reaches at iteration 48. The train and test
performance are close to the validation. Error in the train
data set has a more considerable value than validation and
test data set in the training process, which means the division
algorithm works correctly, and the train data set represents
all data features. Finally, the confusion matrix is presented
in Figure 8.
Figure 8 shows the confusion matrix of the 148 unstable
cases and 152 stable events.All these events are predicted
correctly by the Proposed ANN. Green squares refer to
correct prediction, and red squares refer to incorrect
prediction. From the figure, four parameters can be defined
as follow:
 TP: True Positive, number of correct stable event
prediction to the total stable events, element (0,0)
from the confusion matrix.
 TN: True Negative, number of correct unstable event
prediction to the total unstable events, element (1,1)
from the confusion matrix.
 FP: False Positive, number of incorrect stable event
prediction to the total stable events, element (0,1)
from the confusion matrix.
 FN: False Negative, number of incorrect unstable
event prediction to the total unstable events, element
(1,0) from the confusion matrix.

1376
Mohammad M. Al-Momani et al., International Journal of Emerging Trends in Engineering Research, 9(11), November 2021, 1373 – 1378

The best classification method should give maximum TP and Table 3. Hidden layer 1 to hidden layer 2 weights
TN and minimum FP and FN. Figure 8 shows that the TP,
TN, FP, and FN are 100%, 100%, 0%, and 0% respectively. O/H3 1 2 3 4 5
So this is the best classification algorithm for this problem. 1 -0.54 1.36 -0.43 -1.51 -0.97
2 0.15 1.69 0.36 1.2 -1.31
3 -1.11 0.9 1.36 -0.08 0.97
4. CONCLUSION
4 1.11 -1.49 -0.37 1.52 1.13
5 0.79 -2.16 -0.09 0.74 -0.11
An efficient method is proposed to detect the transient 6 -1.26 -0.39 1.47 0.86 -1
instability based on rotor angle speed obtained through 7 0.1 -1.21 -0.87 -1.35 0.64
synchrophasor measurements across the system in real-time. 8 1.5 -1.46 0.57 -1.01 1.46
The critical feature of the method is that the developed
FFNN output depends upon the post-fault scenarios and is Table 4. Hidden layer 3 to output layer weights
independent of fault location and type of fault. The proposed
ANN used two cascade ANN, one FFNN and the second is H3 1 2 3 4 5 6 7 8
step perceptron. The benefit of usingtwo-level makes a rest O -1.2 -3.9 -1 3.54 2.25 1.35 1.15 4.36
pointinsidethe ANN structure, which helps the learning
algorithm fit the globularity more accurately. The output of REFERENCES
the ANN can be utilized for initiating emergencycontrol
actions. The proposed method is suitable for detecting [1] Odienat, Abdullah, et al. "Low Frequency Oscillation
transient instability and early initiation of the Remedial Analysis for Dynamic Performance of Power
Action Scheme, considering the time requirements for Systems." 2021 12th International Renewable
Engineering Conference (IREC). IEEE, 2021.
implementing the emergency actions.
[2] M. M. Almomani, A. Odienat, S. F. Al-Gharaibeh and
K. Alawasa, "The Impact of Wind Generation on Low
APPENDIX Frequency Oscillation in Power Systems," 2021 IEEE
PES/IAS PowerAfrica, 2021, pp. 1-5, doi:
Table 1. Input hidden layers weights 10.1109/PowerAfrica52236.2021.9543283.
[3] Almomani, Mohammad M., and Seba F. Algharaibeh.
I/H1 1 2 3 4 5
"Modelling and Testing of a Numerical Pilot Distance
1 -0.513 -0.216 -0.113 -0.172 -0.018
Relay for Compensated Transmission Lines."
2 0.125 -0.048 -0.148 -0.176 -0.259
3 -0.3 -0.481 -0.402 0.234 -0.084
International Journal of Scientific Research and
4 0.024 0.422 -0.719 -0.326 -0.608 Engineering Development 3.6 (2020): 776-786.
5 0.138 0.206 0.716 0.038 -0.594 [4] Al-Momani, Mohammad M., Asmaa SM Hatmi, and
6 -0.056 -0.348 -0.413 0.05 -0.471 Seba F. Al-Gharaibeh. "European Journal of
7 -0.467 0.315 0 0.378 0.448 Electrical Engineering." European Journal of Electrical
8 -0.437 0.132 -0.004 0.112 -0.431 Engineering 23.3 (2021): 197-205.
9 -0.319 -0.118 -0.47 -0.109 -0.513 [5] Al-Odienat, A. I., et al. "Connectivity Matrix
10 -0.371 0.509 -0.709 0.092 -0.543 Algorithm: A New Optimal Phasor Measurement
11 0.298 -0.116 -0.051 0.426 -0.322 Unit Placement Algorithm." IOP Conference Series:
12 0.161 -0.213 0.298 -0.362 0.021 Earth and Environmental Science. Vol. 551. No. 1. IOP
13 0.531 0.445 1.021 -0.733 0.531 Publishing, 2020.
14 0.401 -0.663 0.068 -0.359 -0.107 [6] Almomani, Mohammad, Ali S. Al-Dmour, and
15 0.776 0.483 0.344 0.515 0.111 SebaAlgharaibeh. "Application of Artificial
16 0.327 0.406 0.42 -0.522 0.245 Intelligence Techniques for Modeling and Simulation
17 0.49 -0.246 0.518 -0.647 -0.477 of Photovoltaic Arrays."
18 1.054 -0.311 0.521 0.428 -0.261
[7] M. Mahdi and V. I. Genc, "Artificial neural network
19 0.227 0.357 -0.314 -0.003 -0.043
0.221 0.013 0.163 -0.342 0.406
based algorithm for early prediction of transient
20
stability using wide area measurements," in 2017 5th
International Istanbul Smart Grid and Cities Congress
Table 2. Hidden layer 1 to hidden layer 2 weights and Fair (ICSG). IEEE, 2017, pp. 17–21.
[8] S. Suliman and T. A. Rahman, "Artificial immune
H1/H2 1 2 3 4 5
system based machine learning for voltage stability
1 1.472 1.716 1.087 0.013 -1.061 prediction in power system," in 2010 4th International
2 0.655 -1.111 0.766 0.995 -0.58 Power Engineering and Optimization Conference
3 -0.827 1.601 -0.133 -0.059 -1.485 (PEOCO). IEEE, 2010, pp. 53–58.
4 0.383 -0.088 0.66 -0.96 0.92 [9] Multi-Terminal, H. V. D. C. "Mohammad Al-Momani,
5 -1.033 0.645 1.23 1.34 1.118 Seba Al-Gharaibeh1, Hussein. D. Al-Majali2 Mutah
University."

1377
Mohammad M. Al-Momani et al., International Journal of Emerging Trends in Engineering Research, 9(11), November 2021, 1373 – 1378

[10] V. Malbasa, C. Zheng, P.-C. Chen, T. Popovic, and M.

Kezunovic, "Voltage stability prediction using active
machine learning," IEEE Transactions on Smart Grid,
vol. 8, no. 6, pp. 3117–3124, 2017.
[11] M. Mahdi and V. I. Genc, "Artificial neural network
based algorithm for early prediction of transient
stability using wide area measurements," in 2017 5th
International Istanbul Smart Grid and Cities Congress
and Fair (ICSG). IEEE, 2017, pp. 17–21.
[12] S. A. Siddiqui, K. Verma, K. Niazi, and M. Fozdar,
"Artificial neural network based early detection of
real-time transient instability for initiation of
emergency control through wide-area synchrophasor
measurements," in 2015 International Conference on
Computer, Communication and Control (IC4). IEEE,
2015, pp. 1–6.
[13] YADDANAPUDI, KIRAN, and PANNALA KRISHNA
MURTHY. "PERFORMANCE ANALYSIS OF
FUZZY LOGIC CONTROLLED HYBRID ACTIVE
DC FILTER (HADF) FOR 12-PULSE HVDC
CONVERTER." International Journal 9.2 (2021).

1378